Five stories. One idea underneath.

The model stopped being the constraint. It is plateauing on legal reasoning (Maryland’s exam study) and its availability is now a federal decision (the Fable freeze). So the pressure moved up the stack: Perplexity ships a ready-made harness into firms, Deloitte’s clients route work to their own agents, and the scarce resource is the person who can run it all, now priced at $440k and still unfilled.

The government just used the off-switch on two frontier models

The Commerce Department lifted its freeze on Anthropic’s Fable 5 and Mythos 5 on July 1, ending 18 days in which the two most capable models on the market were switched off by federal order over a jailbreak-driven security concern. Access returned after Anthropic agreed to detect risks proactively and report malicious activity to the government.

Issue 009 called the model layer rented and revocable. This is what revocable looks like in practice: a regulator, not a vendor, deciding whether the model under your workflow runs today. If your practice depends on a frontier model, a federal agency now sits in your uptime chain.

[ Al Jazeera · US Lifts Restrictions on Powerful AI Models Fable and Mythos, Anthropic Says ]

AI stopped getting better at law school exams

University of Maryland Carey Law researchers re-ran their law-exam experiment with GPT-5.5 and found no meaningful improvement over last year’s o3 results, even after spending extra inference-time compute. The authors note the apparent plateau matches what other legal benchmarks are showing.

For two years the default legal-AI strategy was to wait for the next model. This is the first credible signal that waiting has stopped paying. If raw legal reasoning has flattened, the gains left on the table belong to the harness and the people running it, not the next release.

[ ABA Journal · Is AI Getting Better at Law School Exams? Maybe Not, Study Shows ]

Perplexity walks into legal through the front door

Perplexity unveiled legal-specific features for its agentic Perplexity Computer platform at an invitation-only New York event: an integration with legal research tool Midpage, connections to document management and contract platforms, and inline citations for cross-jurisdictional research. Gunderson Dettmer has already deployed it enterprise-wide. Pricing was not disclosed.

The consumer-AI giants are no longer selling lawyers an API and a wish. They are shipping the harness itself, pre-wired into the tools firms already run. The fight for the lawyer’s desktop now includes players with hundreds of millions of users behind them.

[ LawSites · Perplexity Makes Its Move Into Legal, Unveiling Industry Features at ‘Computer for Counsel’ Event ]

Deloitte: clients will route 30% of legal work to their own agents

Deloitte’s new legal-industry report projects AI agents handling 30% of corporate legal work within three to five years, with GCs cutting external spend 20-40% and hourly billing falling from 72% to 44% of matters. The number that should worry firms is smaller: only 4% of respondents say they have directly seen benefits from their outside firms’ AI use.

Clients are not waiting for their law firms to modernize. They are building their own agents and re-routing the work. Whatever firms are doing with AI, the buyers say they cannot see it, and invisible efficiency does not protect a fee.

[ Artificial Lawyer · ‘AI Agents Will Handle 30% of Inhouse Work’ – Deloitte ]

Big Law is bidding $440k for AI directors it cannot find

At least 16 major firms, including Latham, Gibson Dunn and Pillsbury, are recruiting AI directors at $200,000 to $440,000 a year and struggling to fill more than 25 open roles. The wish list is 10-plus years of AI experience, and the competition for those candidates is Anthropic, Google and OpenAI themselves.

Here is the bottleneck in one job posting. The models are plateauing, the harnesses are for sale to anyone, and the scarce input is a person who can make the stack work inside a law firm. The market has already repriced that person. Most firms have not.

[ Bloomberg Law · The Hottest Job at Big Law Firms Is Becoming Difficult to Fill ]

Issue 009: the model is rented. Issue 010: the product is the harness. This week the last piece landed. The model is plateauing and Washington holds its switch, so the race is for permission and personnel: the operators who make the stack work inside a firm. Hire accordingly. – The Editors▌

Exhibit AI reports on the AI industry. We do not provide legal advice. Sponsorships are disclosed and never shape coverage. Issue N°011 published MON 6 JUL 2026.